RFR: 8356176: C2 MemorySegment: missing RCE with byteSize() in Loop Exit Check inside the for Expression [v3]

[Emanuel Peter]

On Tue, 12 Aug 2025 09:34:37 GMT, Manuel Hässig <mhaessig at openjdk.org> wrote:

>> A loop of the form
>> 
>> MemorySegment ms = {};
>> for (long i = 0; i < ms.byteSize() / 8L; i++) {
>>     // vectorizable work
>> }
>> 
>> does not vectorize, whereas
>> 
>> MemorySegment ms = {};
>> long size = ms.byteSize();
>> for (long i = 0; i < size / 8L; i++) {
>>     // vectorizable work
>> }
>> 
>> vectorizes. The reason is that the loop with the loop limit lifted manually out of the loop exit check is immediately detected as a counted loop, whereas the other (more intuitive) loop has to be cleaned up a bit, before it is recognized as counted. Tragically, the `LShift` used in the loop exit check gets split through the phi preventing range check elimination, which is why the loop does not get vectorized. Before splitting through the phi, there is a check to prevent splitting `LShift`s modifying the IV of a *counted loop*:
>> 
>> https://github.com/openjdk/jdk/blob/e3f85c961b4c1e5e01aedf3a0f4e1b0e6ff457fd/src/hotspot/share/opto/loopopts.cpp#L1172-L1176
>> 
>> Hence, not detecting the counted loop earlier is the main culprit for the missing vectorization.
>> 
>> So, why is the counted loop not detected? Because the call to `byteSize()` is inside the loop head, and `CiTypeFlow::clone_loop_heads()` duplicates it into the loop body. The loop limit in the cloned loop head is loop variant and thus cannot be detected as a counted loop. The first `ITER_GVN` in `PHASEIDEALLOOP1` will already remove the cloned loop head, enabling counted loop detection in the following iteration, which in turn enables vectorization.
>> 
>> @merykitty also provides an alternative explanation. A node is only split through a phi if that splitting is profitable. While the split looks to be profitable in the example above, it only generates wins on the loop entry edge. This ends up destroying the canonical loop structure and prevents further optimization. Other issues like [JDK-8348096](https://bugs.openjdk.org/browse/JDK-8348096) suffer from the same problem
>> 
>> ## Change Description
>> 
>> Based on @merykitty's reasoning, this PR tracks if wins in `split_through_phi()` are on the loop entry edge or the loop backedge. If there are wins on a loop entry edge, we do not consider the split to be profitable unless there are a lot of wins on the backedge.
>> 
>> <details><summary>Explored Alternatives</summary>
>> 1. Prevent splitting `LShift`s in uncounted loops that have the same shape as a counted loop would have. This fixes this specific issue, but causes potential regressions with uncounted loops.
>> 2. I...
>
> Manuel Hässig has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Improve comment
>  - Fix build failure on product

src/hotspot/share/opto/loopopts.cpp line 233:

> 231: #ifndef PRODUCT
> 232:   if (TraceLoopOpts) {
> 233:     tty->print("Split N%d through Phi N%d in %s N%d", n->_idx, phi->_idx, region->Name(), region->_idx);

Why not also display the `n->Name()`?

test/hotspot/jtreg/compiler/loopopts/superword/TestMemorySegmentByteSizeLongLoopLimit.java line 62:

> 60:     @Test
> 61:     @IR(counts = {IRNode.LOAD_VECTOR_I, IRNode.VECTOR_SIZE_ANY, "> 0",
> 62:                   IRNode.ADD_VI, IRNode.VECTOR_SIZE_ANY,        "> 0",

Oh I only just saw it now. Do you think we can specify the `VECTOR_SIZE` more precisely?
If we don't assert the vector size, it could be that we generate vectors that are smaller, and produce lower performance.

The issue may have been that we have mixed types here.

I wonder if this approach would work for you here:
`test/hotspot/jtreg/compiler/loopopts/superword/TestUnorderedReductionPartialVectorization.java:                  IRNode.VECTOR_CAST_I2L, IRNode.VECTOR_SIZE + "min(max_int, max_long)", "> 0",`

test/hotspot/jtreg/compiler/loopopts/superword/TestMemorySegmentField.java line 57:

> 55:     @Test
> 56:     @IR(counts = {IRNode.LOAD_VECTOR_B, IRNode.VECTOR_SIZE_ANY, "> 0",
> 57:                   IRNode.ADD_VB, IRNode.VECTOR_SIZE_ANY,        "> 0",

Do you actually need the `IRNode.VECTOR_SIZE_ANY` here? Is the default failing for you?
Suggestion:

    @IR(counts = {IRNode.LOAD_VECTOR_B, "> 0",
                  IRNode.ADD_VB, "> 0",

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/26429#discussion_r2269738272
PR Review Comment: https://git.openjdk.org/jdk/pull/26429#discussion_r2269751744
PR Review Comment: https://git.openjdk.org/jdk/pull/26429#discussion_r2269755464